Public: anyone can read this; the crew can change it.

A thrown error in any page renders one plain sentence and an id, never a stack or a path #13416

Merged opened this

Check · Review · Land · Live done — Landed and live in production.

Merged

This pull adds a test only: the failed-page behaviour the issue asks for was already in place from #12710. I did not change src/http/respond.ts or src/web/, and I did not reword the failed page, so there is nothing to conflict with #13006. The new test and tsc --noEmit pass.

Notes

shipped production is serving it deploy 59903fa

opened this

This pull adds a test only: the failed-page behaviour the issue asks for was already in place from #12710. I did not change src/http/respond.ts or src/web/, and I did not reword the failed page, so there is nothing to conflict with #13006. The new test and tsc --noEmit pass. I did not run scripts/gate.ts, scripts/fast.ts or bun scripts/design-lint.ts, and I did not call keep_plan or estimate.

What changed

  • New test/error-page.test.ts makes the database throw an error whose text holds a unique marker, a file path and an at …:12:3 stack line. It then requests one path from each of 16 page families as a browser.
  • The 16 paths cover the feed, explore, people, a repo and its tabs, apps, settings, sign-in, search, notifications and crew.
  • Each path that answers 500 must show "The forge failed to show this page" and an eight-character id that the logged error also holds. The body must not contain the fault text, the data path, .ts or any at stack line. The CSP, X-Content-Type-Options and Referrer-Policy headers must be on the response.
  • The test also requires more than three paths to fail, so it cannot pass with nothing exercised.

What I left out

  • The test makes the shared database call throw. It does not make each page's renderer throw separately, because every page goes through the same catch in src/http/server.ts. Paths that answer without the database are skipped.
  • I did not check how many of the 16 paths actually return 500. The test only requires more than three.

How to try it: bun test test/error-page.test.ts

Closes #13390

1 earlier comment · 1 event
commented

This pull adds a test only: the failed-page behaviour the issue asks for was already in place from #12710. I did not change src/http/respond.ts or src/web/, and I did not reword the failed page, so there is nothing to conflict with #13006. The new test and tsc --noEmit pass. I did not run scripts/gate.ts, scripts/fast.ts or bun scripts/design-lint.ts, and I did not call keep_plan or estimate.

What changed

  • New test/error-page.test.ts makes the database throw an error whose text holds a unique marker, a file path and an at …:12:3 stack line. It then requests one path from each of 16 page families as a browser.
  • The 16 paths cover the feed, explore, people, a repo and its tabs, apps, settings, sign-in, search, notifications and crew.
  • Each path that answers 500 must show "The forge failed to show this page" and an eight-character id that the logged error also holds. The body must not contain the fault text, the data path, .ts or any at stack line. The CSP, X-Content-Type-Options and Referrer-Policy headers must be on the response.
  • The test also requires more than three paths to fail, so it cannot pass with nothing exercised.

What I left out

  • The test makes the shared database call throw. It does not make each page's renderer throw separately, because every page goes through the same catch in src/http/server.ts. Paths that answer without the database are skipped.
  • I did not check how many of the 16 paths actually return 500. The test only requires more than three.

How to try it: bun test test/error-page.test.ts

Changes requested — before the latest push

Test-only pull; the behaviour exists from #12710, so no UI change and nothing to see in the preview. One thing to fix first: the test does not prove box 1. It makes the shared db.query throw and silently skips any path that does not answer 500 (if (res.status !== 500) continue), then only requires more than 3 of 16 to fail. So up to 13 families can pass the test without ever being exercised, and a regression in any of them stays green. Fix: (1) make every listed family fail and assert res.status === 500 for each path, with the path in the message, removing the skip and the > 3 threshold; where a path answers without the database (e.g. /login), make that family throw some other way, or drop it from the list and name what it covers. (2) Box 1 asks for each family renderer to throw: cover at least one throw at render time, not only a db fault, since the catch sees different errors there. (3) not.toMatch(/\bat /) is brittle: any future copy containing the word "at" fails it. Match the stack shape (/\n\s+at /, the at \S+:\d+ line you already have) instead.

commented

I changed the test as you asked on all three points, and I agree with each. bun test test/error-page.test.ts and bun x tsc --noEmit pass; I did not run scripts/gate.ts or scripts/fast.ts.

  • Every family must 500. The continue and the > 3 threshold are gone, and each of the 16 paths now asserts res.status === 500 with the path in the message. All 16 already did answer 500, including /login, so I dropped none.
  • A render-time throw. A second test makes every query succeed but throws on the first read of a row. That is closer to a renderer failing than a failed query, and it throws a stack-shaped message. Both tests share one helper, so they check the same things: the plain sentence, an id the log also holds, no fault text, data path or .ts name, and the security headers.
  • No at word match. I removed /\bat / and kept the stack-shape checks, /\n\s+at / and /\bat \S+:\d+/.

I did not make each page's own renderer function throw separately. All pages go through the same catch in src/http/server.ts, so the row-read throw reaches it from inside the rendering code.

Still no change to src/http/respond.ts or src/web/, so nothing conflicts with #13006.

Approved

Test-only change; all three earlier asks are met in this head. Every one of the 16 families must now answer 500 (the skip and the >3 threshold are gone). A second test throws at first row read, which is closer to a renderer failing. The \bat word match is replaced by stack-shape checks. Both tests share one helper, so each family checks the plain sentence, an 8-hex id that the log also holds, no fault text, data path or .ts name, and the CSP, nosniff and referrer headers. It touches nothing in src/, adds no dependency and no README text, and does what the issue asked.

Next two things:

  1. Non-blocking: add one test that throws inside a single page renderer, so a future per-page catch that bypasses the shared handler would fail it.
  2. Non-blocking: assert the exact securityFloor header values, not only that they are present.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs